Answers are grounded in this paper and cite their sources.
Sixth paper of the human-referential fitting theory program. Between 2024 and 2026, "world model" expanded from a reinforcement-learning textbook term into a high-frequency word shared by industry and the academy, without consensus on what it fundamentally is. This paper extends the program's conceptual toolkit to the term with four contributions. (1) Definition: the world model is a functional, not an entity, concept—a system has one relative to an environment and an action space iff it possesses an action-conditioned, interventional, repeatedly callable prediction capability; implementation splits into written-in, learned, and learn-then-formalize families. (2) Clarification: the confusion is not semantic vagueness but positional polymorphism—one functional concept located at four positions (agent's component, evaluator's benchmark, planner's state representation, trainer's simulated environment), with four forms of invocation; definitional disputes reduce to disputes over positions. (3) Analysis: three propositions—perspective is a property of position; the probability share is a function of the observation position, with calibration as the test standard; environment members divide by the origin of their tasks, and where physics runs out is where agents begin—plus a limit on recursive self-improvement: sustainable self-improvement must retain real-data admixture, and calibration cannot self-certify. (4) Distinctions: from intelligence (policy versus model; acting and predicting do not imply each other) and from AGI (criterial mismatch; component progress is non-transferable), with a structural analysis of the terminal "intelligent world model" in four routes, four requirements, and three boundaries. Eight falsifiability clauses and an admission protocol for new cases close the paper.
No takes yet. Share an insight, caveat, or question.
Yugui Yu (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: